TokenCheat blog

We ran 100 real CLAUDE.md, AGENTS.md, and .cursorrules files through our deterministic optimizer. The median config is fine. The tail is not.

Your Anthropic invoice is one line. Your auditor, your gross margin, and your tax return all need it split. Nobody has the attribution data — yet.

A framework for engineering managers who need to justify AI coding spend to leadership with real numbers.

A 500-line file with 200 lines of dead code means 40% waste on every agent read — here is how to find and fix it.

Every MCP tool definition silently costs 100-400 tokens per turn — here is how to measure and reduce the damage.

Technical breakdown of caching mechanics, pricing discounts, and the mistakes that silently bust your cache.

CLI-first agentic coding vs IDE-native AI assistant — who should use which, and what does it actually cost?

Your CLAUDE.md is prepended to every session — here is how to keep it lean and effective.

Your team spends $500-2K per developer per month on AI coding tools — here is how to manage it like infrastructure, not petty cash.

Context engineering is the discipline of controlling what your AI agent reads — and what it costs you.

Same repo, different packing strategies — what actually changes token load and retrieval quality.

A snapshot of flagship and efficiency tiers for coding workloads — pricing moves fast.

When cache reads beat full input pricing — and when they don't.

Tool calls are not free — they are recurring input tokens with JSON-shaped surprises.

Practical levers: cache ratio, tool surface, and model routing — without dumbing down agents.